Multistep temperature prediction for proactive thermal management on chip multiprocessors

In chip multiprocessors, the thermal hot spot prediction is vital to take proactive thermal management decisions for mitigating its impact on the performance of the application. Core temperature prediction for a range of future time steps is challenging because workload changes dynamically at runtim...

Full description

Saved in:
Bibliographic Details
Published in:The Journal of supercomputing 2021-08, Vol.77 (8), p.8967-8994
Main Authors: Revathi, N., Sumathi, G.
Format: Article
Language:English
Subjects:
Algorithms
Compilers
Computer Science
Interpreters
Iterative methods
Multiprocessing
Process migration
Processor Architectures
Programming Languages
Thermal management
Throttling
Workload
Workloads
Citations: Items that this one cites
Items that cite this one
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:In chip multiprocessors, the thermal hot spot prediction is vital to take proactive thermal management decisions for mitigating its impact on the performance of the application. Core temperature prediction for a range of future time steps is challenging because workload changes dynamically at runtime. This article presents an adaptive multistep temperature prediction (MSTP) algorithm, which uses iterative steps to generate core temperature predictions for a range of future time steps using the initial temperature information and the application workload profile obtained offline. MSTP considers the autocorrelation of temperature and addresses the nonlinear heteroscedasticity across observations. We evaluate the proposed MSTP on a real hardware platform and compare it to the model-based approach for a wide range of parallel benchmarks. Experimental results show that the proposed algorithm achieves more than 96% prediction accuracy and minimizes the heteroscedastic residual variance. In case of thermal hot spot prediction, we show that the performance throttling problem is alleviated using a process migration mechanism with a minimal migration overhead.
ISSN:0920-8542
1573-0484
DOI:10.1007/s11227-020-03611-5